Sleep Stage Classification from Wearable Sensor Data: Toward Transparent and Accessible Health AI
Date:
Abstract
Sleep quality plays a vital role in cognitive functioning, physical health, and emotional well-being, yet most people lack access to accurate and effective sleep monitoring. While clinical tools like polysomnography offer detailed insights, they are expensive and impractical for continuous or at- home use. This research explores whether consumer-grade wearable devices, as part of broader cyber-physical systems for health monitoring, can be used to classify sleep stages accurately and responsibly using fewer and more accessible physiological features.
We use the DREAMT dataset, which contains wrist-based physiological signals from Empatica E4 devices alongside expert-labeled sleep stages for 100 participants. We trained two deep learning architectures, Long Short-Term Memory (LSTM) and Convolutional Neural Networks (CNN), on both full-feature and reduced-feature datasets derived from wrist-worn physiological signals. The reduced set, which includes only heart rate, inter-beat interval, and triaxial accelerometry, reflects signals commonly available on consumer-grade wearables.
Despite this limited input, our reduced-feature LSTM model achieved 92% overall accuracy, demonstrating strong performance while maximizing generalizability. CNNs showed comparable results with fewer parameters, making them attractive for deployment on low-power devices. To counter the severe under- representation of light-sleep (N1) and deep- sleep (N3) epochs, we paired focal loss with stage-specific class weights, steering the models’ learning toward these minority classes and preserving overall balance across all sleep stages.
This work contributes to a broader open-source effort to empower users to see, understand, and control the data collected from their wearables. Our approach enables users to trace exactly which features were used, which algorithms were applied, and how confident the model is in its predictions. This study shows how transparent AI models can power personalized and responsible sleep health tools outside of clinical settings